VCSEL Light Coupling to a Waveguide to Interconnect XPUs and HBMs on Interposer Chips

Abstract

Multimode VCSELs coupled into waveguides can be a practical path toward realizing commercially viable photonic interposer chips. The experimental coupling of multimode VCSEL light to a non-silicon waveguide fabricated using a CMOS-compatible process is demonstrated. The GaP prism was tested and adopted as a coupling method. Both conventional and cavity-type optical waveguides, fabricated from CMOS-compatible PECVD SiO₂, Si₃N₄, and SiO_xN_y materials, were evaluated. The average propagation loss transmitted through the cavity-type waveguide was measured as 0.444 dB/cm. A polyimide micro-lens, cavity-type waveguide, and a GaP prism coupler are developed to inject the multimode VCSEL light into the waveguide measuring the net coupling loss of 0.762 dB. The packaged size of VCSEL has an area of 0.4 mm² and a height of 0.64 mm. MUX/DeMUX was designed on the bottom of the prism. A light source, a modulator, and MUX/DeMUX are all located in the same area of the prism bottom in VCSEL-based interconnections.

1. Introduction

Since the integration of GPUs with high-bandwidth memory (HBM) on silicon interposer chips using 2.5D packaging technology in 2015 [1], GPU and AI chips equipped with HBM have led the semiconductor chip market. These high-performance chips have widely been deployed in data center servers. As the performance of individual chips has been enhanced by advancements in technologies such as sub-2 nm nodes and gate-all-around (GAA) transistors, the communication bandwidth between switch chips and servers also needs to be improved to keep pace with the performance of individual chips.

While the electrical signal transmission rates per lane for HBM3E and HBM4 are 9.6 Gb/s and 6.4 Gb/s, respectively, according to JEDEC, the optical transmission rate of 106.25 Gb/s per channel was standardized in 2022 by IEEE 802.3 db, and a 200 Gb/s standard is currently underway [2,3]. Optical signals can offer a tenfold increase in communication bandwidth per channel compared to electrical signals.

Since the silicon photonics began attracting a lot of interest in the photonics research community in the early 2000s, many optical communication-based network architectures for CPU applications have been proposed, including a 1000-core on-chip optical network [4], 256 cores with 1024 threads [5], point-to-point optical networking on 8 × 8 macrochips [6], and an optical network based on ring resonator switching elements [7]. However, these architectures have not yet reached the stage of experimental integration and testing on actual chips. The difficulty can be largely attributed to the properties of silicon used as an optical waveguide material. Light sources, modulators, MUX/DeMUX devices, and coupling components designed for single-mode silicon waveguides are still not sufficiently advanced to realize the optical network-on-chip proposed by these architectures. Since the mid-2010s, the research focus has shifted to co-packaged transceivers, which interconnect switches to servers or to other switches in data centers using silicon photonics-based technologies [8]. In recent years, advancements in light sources for co-packaged optics (CPO) and silicon photonics have been reported [9,10]. As research funding from large companies and governments has decreased, many startups aiming to supply optical devices integrated into silicon wafers or chips have been founded.

Table 1. CPO technology trend.

	Pluggable on Board 2024	CPO on Substrate 2026	CPO on Chip 202x
Bandwidth	1.6 Tb/s	6.4 Tb/s	12.8 Tb/s
Power	1×	<0.5×	<0.1×
Latency	1×	<0.1×	<0.05×

lists the CPO technology trend predicted within the photonics industry [11]. As of 2024, optical transceivers are plugged in at the server board edge. By 2026, they are expected to be co-packaged with XPU or switch chips on the substrate and eventually co-packaged on the silicon chip. Currently, a single optical transceiver supports a data rate of 1.6 Tb/s. It is expected to achieve a tenfold increase in bandwidth, a tenfold reduction in power consumption per gigabit, and a twentyfold decrease in transmission latency on the silicon chip.

Through wavelength division multiplexing (WDM), a single optical waveguide can carry eight different wavelengths [12], enabling over 160 times the bandwidth capacity compared to a single electrical lane. Based on a 12.8 Tb/s data rate, electrical signals would require 2560 wires (1280 lanes), while optical signals using two wavelengths per channel (64 channels total) and four channels per waveguide would require only 16 optical waveguides. This clearly indicates that, without adopting optical signaling, the required number of electrical interconnects will soon exceed what can be physically integrated within an interposer chip [13]. Considering the typical dimensions of a VCSEL and a photodiode (250 µm × 250 µm in area and 150 µm in height), 64 channels require 64 VCSELs and 64 photodiodes, corresponding to a total area of approximately 8 mm². This is much smaller than the area of a typical processor chip (>100 mm²). Since a prism coupler can be fabricated as small as a VCSEL, the area required for photonic chips should not pose a limitation.

In data center applications, the majority of communication links between servers and switches, or between switches, are shorter than 100 m. Copper cables are typically used for links spanning several meters, whereas multimode fibers are employed for distances beyond the reach of copper cables. Interconnections using multimode VCSELs are among the most effective optical solutions for photonic interposer chips, where link lengths are typically a few tens of centimeters. The coupling technology of multimode VCSELs to optical waveguides on a silicon wafer is critical for realizing a commercial photonic interposer chip.

The technology of VCSELs has advanced rapidly. From the mid-2010s, 50 Gb/s products became commercialized and utilized in supercomputer production. Directly modulated multimode VCSELs achieving data rates of 200 Gb/s per lane have been reported, including 224 Gb/s [14], 200Gb/s PAM 4 [15], and 100 GBd PAM 4 [2]. Four-channel CWDM transmissions using optical signals of PAM4 (4 × 112 Gb/s) [16], NRZ (4 × 50 Gb/s) per fiber [17] and VCSEL-based co-packaged transceivers [18] have been developed, including NRZ 1.3 pJ/b [12] and novel packaging structures [19,20]. Single transverse mode VCSELs have also been developed, demonstrating a 3 dB direct modulation bandwidth of 39.4 GHz, an output power of 2.8 mW at 8 mA driving current, a 50 Gb/s NRZ signal transmission over 2 km [21], a high output power of up to 16 mW [22], and 16 channels with 70 Gb/s PAM4 per channel transmitted over 5 km [23]. VCSELs and their properties can be designed and predicted by using machine learning-based transformer neural networks [24].

Several studies have reported the coupling of VCSEL light into optical waveguides reporting the insertion loss of −11.8 dB [25]. All papers have focused on coupling single-mode VCSEL light into single-mode silicon waveguides [26,27]. However, none of the studies employed microlenses, resulting in excessive coupling losses greater than 10 dB due to the modal size mismatch between the VCSEL output beam and the single-mode silicon waveguide.

In this study, we developed a polyimide microlens, a cavity-type waveguide, a GaP prism coupler, and designed a thin-film MUX/DeMUX. For the first time, to the best of our knowledge, we demonstrated the injection of multimode VCSEL light into a non-silicon waveguide fabricated using a CMOS-compatible process. We achieved sufficiently low coupling and propagation losses to make this approach feasible for real commercial photonic interposer chip products. A primary advantage is the compact footprint, where the light source, modulator, and MUX/DeMUX are all located in the same area of the prism bottom in VCSEL-based interconnections.

2. Experiments

PECVD Si₃N₄/SiO_xN_y optical waveguides have been studied by several groups, demonstrating the feasibility of integrating SiON waveguides with silicon waveguides on SOI [28], reporting the propagation loss of 0.2–0.3 dB/cm of the waveguide [29], examining the influence of hydrogen on the propagation losses [30], and inventing a coupling method [31]. Among the methods that can inject light into PECVD Si₃N₄/SiO_xN_y optical waveguides, butt coupling and grating coupling, which are widely used in silicon photonics, are difficult to apply. Given the size of the VCSEL’s top surface, facet-to-facet contact is feasible only at the edge of the chip, where optical waveguides are formed on the photonic interposer. As a result, signals from a directly modulated VCSEL can be generated only at the chip edges, making it impossible to generate such signals at locations away from the periphery. It is also difficult to secure space to attach them to the side of the chip. For grating coupling, the difference in refractive index between silicon oxynitride and silicon dioxide is small (Δn = 0.1 − 0.55), which leads to a low coupling efficiency. Excluding the two methods, the only possible method is to use a prism. To use the prism method, the following three technologies must be developed.

First, an adhesive technology that is transparent, satisfies the refractive index matching, maintains its adhesive strength at 200–400 °C, and is not affected by solvents used in the CMOS cleansing process such as BOE, isopropanol, acetone, etc., needs to be achieved. There are no commercial adhesives that are transparent with a refractive index greater than 1.6 for a wavelength of 850–1550 nm.

Second, a microlens that can produce collimated light under the conditions shown in Figure 1a must be developed. Microlenses made of quartz and silicon are commercially available, but their refractive indices (n₁ = 1.45 for quartz and n₁ = 3.45 for silicon) are not suitable for the current application. When a commercial adhesive with a refractive index of n₂ = 1.4~1.5 is used, the focal length calculated by the equation f = R·n₁/(n₁ − n₂), where R is the microlens radius (15 μm), becomes either too long (>435 μm for quartz) or too short (<26.6 μm for silicon) for practical use. The VCSELs fabricated with such microlenses in references [32,33,34] may not match the refractive index for our purpose and are not expected to be commercially available.

Third, a special packaging technology for VCSEL and PD needs to be developed. The current commercial package has a hermetically sealed structure filled with nitrogen gas, which maintains a specific distance between the laser light exit and the lens. This type of packaging is overly bulky, and dozens to hundreds of VCSELs and PDs cannot be integrated onto a chip. We developed an adhesive technology, a stand-alone polyimide microlens, and a packaging technology that satisfy these requirements.

Figure 1b shows a micrograph of a GaP prism bonded to the surface of a SiON optical waveguide and (c) shows an enlarged photograph of the side view. The GaP prism was fabricated by grinding and polishing a piece of GaP wafer that was attached to an aluminum block. The inclined surface angle of 26.5° was precisely transferred from the accurately machined aluminum block, which also had an inclined surface of 26.5°. Although the prism can be made as small as a typical VCSEL, for experimental convenience, its width and height in Figure 1b are in the range of 1–2 mm.

Thermoset polyimide and our specially developed procedure were used for bonding GaP prisms to the silicon wafer on which SiON optical waveguides were formed [35,36]. The process used to induce adhesion on the imidized thermoset polyimide layer is similar to that used for oxide layers in the fabrication of SOI wafers. First, dangling bonds are generated on one of the imidized polyimide surfaces through grinding and polishing. After coating the silicon wafer with liquid polyimide and carrying out the imidization process, the wafer was ground and polished using a wafer polishing machine. For the prism, liquid polyimide was also applied and imidized, but without grinding and polishing. We found that the presence of dangling bonds on only one surface was sufficient to induce adhesion strong enough for our application. Then, the two surfaces are bonded by applying pressure at a temperature between 120 °C and 150 °C. We believe that adhesion is achieved through the rejoining of dangling bonds between the two thermoset polyimide layers coated on the silicon wafer and the prism bottom, which are otherwise non-adhesive under normal conditions. The adhesion characteristics between the prism and the wafer are shown in

Table 2. Specification.

Refractive index	1.69~1.72 at 850 nm
Transmission	>99% for 2 μm thickness at 850 nm
High temperature resistance	400 °C (compatible with CMOS metal process)
Solvent resistance	BOE, acetone, isopropanol, etc., (unaffected by alcoholic cleaning)
Adhesion strength	>1 N/mm2 (not detachable before breaking GaP prism at 25 °C)

. It had a transmission of over 99% at a thickness of 2 μm required for adhesion, and the refractive index was measured to be up to 1.72 depending on the heat treatments applied to the polyimide. When force was applied from the side to observe the adhesion strength, the prism broke more frequently than it fell off.

The radiation angle of VCSEL in air is typically in the range of 20–40°. The microlens must have the right value of the refractive index to collimate the light emitted by the median value of 30°. Figure 2 shows a scanning electron microscope (SEM) image of the polyimide microlens array (a) and an enlarged image (b) of one. The same thermoset polyimide used to attach the prism to the wafer was also employed for microlens fabrication. We believe that a suitable process can be established with any brand of thermoset polyimide, as long as it has a similar imidization temperature to those listed in

Table 3. Properties of microlens.

Properties	Value
Refractive index at 850 nm	1.72
Transmission loss	~1.8%
Diameter	30 μm
Material	Polyimide
Imidized temperature	330 °C
Coefficient of thermal expansion	38 ppm

. The principle of forming a hemispherical shape from liquid-state polyimide is well known and has been reported in previous studies [32,33,34]. A cylindrical shape can be transformed into a hemispherical lens by heating and liquefying, a self-driven process caused by the surface tension of the liquid. We fabricated cylindrical polyimide structures on a silicon wafer using a properly designed mask and photolithography. The initial cylinder diameter was set to 40 μm in order to obtain microlenses with a final diameter of 30 μm, as the size is reduced during multiple heating steps. Based on the curve profile, the final shape of the microlens falls between a paraboloid and a hemisphere, which may be attributed to the thick base of the prepared polyimide sample. Since a parabolic lens is free from spherical aberration, our microlens is expected to exhibit better aberration characteristics than a perfect spherical lens. The properties of the polyimide microlens are summarized in Table 3.

After attaching the microlens to the 25 Gbps VCSEL chip, the shape of the output light was observed as shown in Figure 2c,d. The picture was taken by placing an infrared card 7 cm vertically from the VCSEL. Since the distance from the VCSEL to the optical waveguide is <1 mm, the collimated light proves the sufficient performance of the lens.

The test results of the microlens were used to complete the package. Figure 3a shows the package structure. The lens was fabricated separately and attached to the VCSEL in the final step. A commercial adhesive with refractive index n₂ = 1.493 and transparency >99% per 100 μm was used. The focal length of the microlens was calculated as 113.65 μm within the microlens material using the equation f = R·n₁/(n₁ − n₂). If the thickness of the flat region of the microlens is 50 μm, then the thickness of the adhesive between the VCSEL and the microlens is calculated to be 42 μm using simple geometry, which is automatically determined during the attachment by optimizing the shape of the collimated beam.

Figure 3b shows a micrograph of the VCSEL packaged with this structure. It has an area of 800 μm × 500 μm and a height of 640 μm. It does not exhibit the minimum size because it was designed for the experimental verification of optical coupling. Considering the net area of the VCSEL, 250 μm × 250 μm, the minimum area could be less than 0.1 mm².

As shown in Figure 4a, the prisms were attached to both ends of the optical waveguide, and the packaged VCSEL was attached to the prism on one side. An experiment was conducted to measure the intensity of VCSEL light entering the optical waveguide and exiting the opposite prism. Figure 4b shows a micrograph of the laser light leaving the opposite prism after it enters the optical waveguide. Figure 4c shows a micrograph of light on the opposite prism, where the camera was off angled from the output light to reduce the brightness.

The alignment and attachment procedures were as follows. The packaged VCSEL was mounted on a holder equipped with electrodes for current injection during alignment. The holder was designed to move along the XYZ axes and to tilt the incident angle of light into the prism. Alignment was optimized by measuring the output power emitted through the opposite prism in real time. Once optimized, the VCSEL was fixed to the prism using a transparent epoxy. The alignment process itself was not time-consuming because the tolerances for a 10% decrease from the peak output power were greater than several micrometers. The initial challenge was to find an approximate incident angle that allowed detectable light through the opposite prism. However, once the optimal angle was found, the alignments for incident angles were straightforward in subsequent trials.

Unlike the losses in references [28,29], the propagation loss of our SiON waveguide was excessively large, being >10 dB/cm for 850 nm multimode light. We attribute the large loss to the irregular distribution of the refractive index caused by a non-uniform distribution of O and N atoms, which leads to a large Rayleigh scattering loss.

We designed a cavity-type waveguide consisting of an oxide spacer and two distributed Bragg reflectors (DBRs) on a silicon wafer, where the DBRs encompass the spacer in two dimensions. The spacer and DBRs were deposited using a CMOS-compatible process with PECVD SiO₂ and Si₃N₄ for the spacer and top DBR, and LPCVD SiO₂ and Si₃N₄ for the bottom DBR. We tested two different structures of cavity-type waveguides. Figure 5a shows the SEM image of the first structure, three periods of DBR/2 μm thick SiO₂ smoother/4 μm thick SiO₂ spacer/four periods of DBR, where each period of DBR has 150 nm thick Si₃N₄ and 642 nm thick SiO₂ layers. After the spacer layer was deposited and dry-etched to a 4 μm target thickness, a 2 μm thick smoother layer was deposited to remove the roughness on the spacer wall. The first structure provides information about the propagation loss caused by light leakage through the top and bottom DBRs while refraining from the loss caused by sidewall roughness. The second structure in Figure 5b is exactly the same as the first except that there is no smoother layer, which provides information about the loss caused by the sidewall roughness.

The propagation loss was measured by comparing the output powers detected from the cleaved edges of 5.5 mm and 10.5 mm long waveguides for two structures as shown in Figure 6a. The light was input into the waveguide through a prism and collected by a 200 μm diameter commercial lensed detector at the edge of the waveguide. Figure 6c shows the micrograph of the detector.

Table 4. The measured data.

Structure 1 (Unit: mW)		Structure 2 (Unit: mW)
5.5 mm	10.5 mm	5.5 mm	10.5 mm
0.176	0.170	0.09	0.08
0.159	0.152	0.117	0.045
0.182	0.170	0.109	0.068
0.169	0.161	0.150	n/a
0.177	0.167	n/a	n/a
Av. 0.1726	Av. 0.164	Av. 0.117	Av. 0.065

shows the measured data for samples 3–5. All data represent measurements obtained from different samples corresponding to each number. Measurements were repeated three to five times for each sample, and the values shown in the table represent either the average or the most representative value among the repeated measurements in order to minimize errors. Raw data recorded in mV were transformed into mW, as the power of VCSEL, which was measured to be 1150 mV by placing the detector 1 mm apart from the VCSEL aperture, then transformed into 1 mW, which represents not absolute power but relative values to compare measurements in mW. The average propagation losses were 0.444 dB/cm for structure 1 and 5.1 dB/cm for structure 2. While the loss caused by the side wall roughness is quite large, the leakage losses through the top and bottom mirrors are sufficiently small to be used for the network-on-chip. The total loss from the VCSEL to the detector for a 10.5 mm long waveguide was 7.85 dB for structure 1.

The output powers emitted at the cleaved edge and through the prism for the same length of the waveguide were measured as shown in Figure 6a,b.

Table 5. The measured data.

Edge (mW)	0.011	0.014	0.013	0.009	0.013	Av. 0.012
Prism (mW)	0.015	0.012	0.017	0.01	0.009	Av. 0.013

shows the measured data. The average value of the prism is 8% higher than that of the edge. While the light emitted at the edge has a large angle due to the small thickness of the waveguide core, the light emitted through the prism has a smaller angle due to the wide area of light at the surface of the waveguide core, leading to better collection efficiency with our detector, which is better than making up for the reflection loss on the two surfaces of the prism. Waveguide widths of 30 μm and 50 μm were designed and tested for multimode VCSEL. The waveguide with a 50 μm width was used for the measurement. One remarkable fact is that the light unaffected by the prism in Figure 6b is negligibly small, being less than 0.1%.

The average coupling loss can be obtained from the data in Table 4 and Table 5. For approximation, we take the same ratio of collection efficiencies 12/13 between the edge and the prism for the measurement in Table 4. If detected through the prism instead of the edge, the output powers for the 10.5 mm long waveguide of structure 1 in Table 4 will be (0.184, 0.165, 0.184, 0.174, 0.181, Av. 0.178), which were derived by multiplying 13/12 to the values in the second column. The total average loss will be 7.45 dB. As the reflection loss of 5.46 dB for the two prisms and the propagation loss of 0.466 dB for the 10.5 mm long structure 1 waveguide are subtracted from the total loss, the average net value of the coupling loss per prism results in 0.762 dB. The loss of 0.466 dB was calculated from the average propagation loss of 0.444 dB/cm and the waveguide length of 10.5 mm (0.444 dB/cm × 10.5 mm). The theoretical value of reflection loss for one prism attached with optical glue is 2.73 dB as calculated in

Table 6. Reflections at interfaces.

Surfaces	Polyimide Adhesive		Optical Glue
	TE	TM	TE	TM
Air/Prism (Glue/Prism)	0.27/0.026 (0.125/0.002)	0.27/0.026 (0.125/0.002)	0.27/0.026 (0.125/0.002)	0.27/0.026 (0.125/0.002)
Prism/Adhe.	0.23/0.0001	0.007/0.053	0.461/0.005	0.008/0.085
Adhe./WG	0.24	0.124	0.054	0.038
Sub-total	0.573/0.26	0.365/0.192	0.628/0.083	0.303/0.143
Total (Glue/Prism)	0.469/0.226 (2.75/1.11 dB) (0.363/0.207) (1.96/1.0 dB)		0.466/0.113 (2.73/0.52 dB) (0.36/0.091) (1.938/0.414 dB)

.

The thicknesses of the DBR layers on the sidewall in Figure 5a,b are different from the target values because the deposition rate on the sidewall is lower than that on the flat surface, causing a large bending loss. Using the CMP process, the DBR on the sidewall can be deposited separately from the top DBR by removing it only on the top surface before the top DBR is deposited, which may produce the designed thickness of the sidewall mirrors.

3. Modeling MUX/DeMUX and AR-Coating

A WDM filter with a verified performance among commercially available products is a thin-film filter in which two thin films with different refractive indices are alternately stacked. Figure 7a shows a schematic diagram of four optical wavelengths transmitted through a single optical waveguide by coating WDM thin-film filters on the bottom of the prisms. For example, λ₃ passes through the prism coated with the filter of the corresponding wavelength and enters the waveguide. It is reflected off the bottoms of the prisms coated with other wavelength filters, proceeds along the waveguide, and is emitted from the prism of the corresponding wavelength. Figure 7b shows the structure of the WDM filter and the light path.

Figure 8 shows reflection and transmission graphs derived from theoretical calculations using TiO₂ and Ta₂O₅ films, which are often used to design WDM filters. The refractive index of the prism is n_prism = 3.16 for 850 nm, that of the TiO₂ is n_H = 2.5086, and that of Ta₂O₅ is n_L = 2.0908. When the prism angle is θ, the light that enters perpendicular to the inclined plane enters the bottom as θ and proceeds as ф, as shown in Figure 7b. The angle θ = 26.5° was used for modeling. The thickness of each film was determined such that the path at the corresponding angle was λ/4. The TiO₂ film was denoted as T_H and the Ta₂O₅ film was denoted as T_L.

Figure 8a,b shows the reflection spectra calculated when the T_H and T_L films were repeated 15 times [(T_H T_L)¹⁵]. Figure 8c,d shows the transmission spectra calculated when a spacer with a width of 4 λ was inserted between the two reflecting mirrors formed by repeating the T_H and T_L films 15 times [(T_H T_L)¹⁵(T_H T_H)⁸(T_L T_H)¹⁵]. Such calculations are technically mature and are used to produce commercial products. The transfer matrices for T_H and T_L layers can be found in optics textbooks. We calculated and plotted the matrices [(T_H T_L)¹⁵] and [(T_H T_L)¹⁵(T_H T_H)⁸(T_L T_H)¹⁵] using the Mathcad program and the parameters n_prism = 3.16, n_H = 2.5086, n_L = 2.0908, n_adhesive = 1.72, and θ = 26.5°. It was assumed that light enters the prism perpendicularly to the inclined plane. Full widths at half maximum (FWHMs) of stopbands are 184 nm and 67 nm. FWHMs of transmission curves are 0.17 nm and 6.32 nm. As shown in the figures, the spectral curves of TM mode are less effective for a WDM filter than TE mode, because reflectivity of TM mode at the interfaces is lower than that of TE mode. A remarkable fact is that both TE and TM modes are allowed to be utilized in our schemes of prism coupling and thin film WDM filters. Other schemes such as grating coupling and AWG filters generally do not allow both TE and TM mode operations simultaneously, which introduces a large complexity of system or a large insertion loss.

Figure 9 shows a schematic diagram of a light path that passes through the two surfaces of the prism and the adhesive and enters the waveguide. There are three interfaces that reflect light from air to the prism, prism to adhesive, and adhesive to the waveguide core. For the adhesive, if heat and cleaning conditions are allowed, a commercial optical glue (n = 1.51) may also be used because it satisfies the refractive index matching between the prism and the cavity-type waveguide. In general, it does not satisfy the refractive index matching between the prism and the SiON waveguide. For the measurements, the data in Table 4 and Table 5 were obtained with prisms attached using an optical glue because this attachment process is fast and inexpensive. To the best of our knowledge, no difference in the coupling efficiencies was observed for the two adhesives when there was no anti-reflection (AR) coating on the prism.

An example of the AR coating layers and thicknesses on the prism is shown in Figure 9. One layer on the inclined plane and two layers on the bottom were deposited with TiO₂ and Ta₂O₅. The theoretical values of the reflections at the three interfaces are summarized in Table 6. The two numbers in each box represent reflections before and after AR coating. The sub-total represents the total reflection loss accumulated by the three interfaces for the TE and TM modes, whereas the total represents the reflection loss calculated from the sub-total losses, 0.5 × TE + 0.5 × TM, to consider the unpolarized VCSEL light. The same transfer matrix method, parameters, and the Mathcad program used for modeling WDM filters were used for modeling AR coating.

When the polyimide adhesive is used, the total reflection loss of 2.75 dB reduces to 1.11 dB after AR-coating. When optical glue is used, the loss of 2.73 dB reduces to 0.52 dB, which is less than half the value of the polyimide adhesive. When optical glue is used for attaching LD to the prism instead of air, the total reflection losses are calculated to be 1.96 dB/1.0 dB and 1.938 dB/0.414 dB, suggesting that the theoretical limit to reduce the reflection loss is near 0.4 dB. The absorption loss through the prism can be ignored because the absorption coefficient of GaP for 850 nm, α = 0.0000/cm, is negligibly small [37].

4. Discussion

The primary advantage of VCSELs is that their light is emitted perpendicular to the wafer surface, enabling laser characterization at the water level without the need to dice the wafer into individual chips. This advantage similarly applies to the fabrication of photonic interposer chips utilizing VCSELs or specially arranged grating couplers for edge-emitting lasers [8]. Figure 4 illustrates transmission experiments of VCSEL signals through optical waveguides. Figure 4a schematically depicts the configuration for coupling VCSEL signals into and out of the optical waveguide via a prism. Figure 4c shows a photograph capturing laser light from the prism surface. Our experiment demonstrates that the coupling and transmission losses are acceptable for application to photonic interposer chips.

Photonic interposer chips employing VCSELs are highly advantageous for mass production, as optical testing can be performed at the wafer level. The process begins by permanently bonding prisms at the predetermined positions. Functional VCSELs are then selected and mounted onto holder equipment. This holder is equipped with electrodes that allow current injection and real-time observation of the emitted laser light. The VCSELs are carefully aligned to their target positions, while simultaneously monitoring light emission through the opposite prisms in real time. Once optimal alignments are achieved, the lasers are temporarily bonded using a transparent adhesive that liquefies at approximately 120 °C.

For high-volume manufacturing, dozens to hundreds of components can be bonded simultaneously, with defective units replaced or reworked as needed. Since the positions of all prisms, VCSELs, and photodiodes on the photonic interposer chip are predetermined, multiple mask chips are prepared to pre-align these components at their designated locations. Once the actual interposer chip is ready, the pre-aligned components can be transferred and bonded in a single step using holding equipment. Even after temporary bonding, the condition of the VCSELs and photodiodes can be verified, and any defective unit can be selectively replaced by reheating the specific area to 120 °C. Once all VCSELs and photodiodes are confirmed to be functional, permanent bonding is performed during the RDL molding process.

Several research groups have reported experimental results on coupling a single mode VCSEL light into and out of single mode silicon optical waveguides [25,26,27]. Due to the high refractive index of silicon (n ≈ 3.45), prisms cannot be used for coupling into the silicon optical waveguide, and diffraction gratings must be employed instead. VCSELs emit unpolarized light, whereas diffraction gratings efficiently diffract only the TE mode at appropriate angles, while the TM mode is not effectively diffracted. As a result, an inherent coupling loss of approximately 3 dB occurs. The mode size mismatch between the laser beam and the waveguide further exacerbates the coupling loss without microlens, because commercial microlenses cannot be used due to their excessive size. Consequently, coupling losses in these studies exceed 10 dB.

Silicon optical waveguides have the drawback of relatively high propagation losses, typically ranging from 1 to 3 dB/cm. Additionally, they require the use of silicon-on-insulator (SOI) wafers. Since the waveguides share the silicon layer with various logic components, this may introduce certain constraints in the overall circuit design. It is more advantageous to form optical waveguides using non-silicon materials such as SiO₂ or Si₃N₄. which are fabricated using CMOS processes in silicon fabs. To the best of our knowledge, this is the first report to demonstrate the injection of multimode VCSEL light into a non-silicon waveguide fabricated using a CMOS-compatible process.

When considering factors such as footprint, power consumption, fabrication cost, and wafer-level testing of optical devices, VCSELs are among the most advantageous options for implementing photonic interposer chips. Nevertheless, several challenges remain. First, it must be demonstrated whether the processes developed in laboratory for fabricating polyimide microlenses and packaging VCSELs and PDs are suitable for mass production. The alignment and attachment of prisms, VCSELs, and PDs to waveguides must also be compatible with high-throughput manufacturing.

Second, for a communication bandwidth of 12.8 Tb/s in Table 1, at least 64 VCSELs and 64 photodiodes are required, with each device supporting 200 Gb/s. Based on current prices, the cost of these optical components is much higher than those of GPU and HBM integrated on a silicon interposer chip. Therefore, it is reasonable to expect that electrical interconnects will remain in use as long as they can meet communication requirements, and that optical interconnects will be introduced progressively over unexpectedly long years. However, given the relatively low manufacturing cost of VCSELs and photodiodes, the total price of optical components is expected to drop significantly once large-scale demand emerges. In such a case, the market size of photonic interposer chips could become comparable to that of HBM or/and GPU.

One drawback of VCSEL-based interconnections is modal dispersion between TE and TM modes, which can occur in cavity-type waveguides due to asymmetry between the horizontal and vertical directions. Future work will analyze, both experimentally and theoretically, the impact of modal dispersion on high-speed transmission over waveguides of several tens of centimeters. If found detrimental, modal dispersion can be avoided by designing waveguides symmetrically by tapering the waveguide width in the horizontal direction.

5. Conclusions

Multimode VCSEL light was coupled to a cavity-type waveguide fabricated using a CMOS-compatible process and materials, SiO₂ and Si₃N₄. The average propagation loss transmitted through the cavity-type waveguide was measured as 0.444 dB/cm. Except for the reflection loss at the prism, the net coupling loss from the VCSEL to the waveguide was 0.762 dB. The packaged size of the VCSEL or photodiode has an area of 0.4 mm² and height of 0.64 mm, which can be attached to the prism.

The prism used to inject light into the waveguide can also be used as a WDM filter in which two thin films with different refractive indices are alternately stacked on the bottom. The theoretical calculations confirmed the performance of the WDM filter. If anti-reflection films were coated on the two surfaces of the prism with materials, TiO₂ and Ta₂O₅, then the reflection loss at the prism was calculated to decrease from 2.75 dB to 1.11 dB for polyimide adhesive and from 2.73 dB to 0.52 dB for optical glue. It was demonstrated that the light source, modulator, and MUX/DeMUX can all located in the same area of the prism bottom in VCSEL-based interconnections. Considering that a prism coupler can be made as small as VCSEL, 250 µm × 250 µm in area and 150 µm in height, the area required for photonic chips should not be limited to realize the commercial photonic interposer chips.